Sen. Maggie Hassan, who is set to assume the chairmanship of the Senate Homeland Security and Governmental Affairs Committee later this year, is preparing to hold legislative hearings focused on artificial intelligence safety. The move follows a new Government Accountability Office (GAO) report indicating that four of the most widely used AI platforms—ChatGPT, Claude, Gemini, and Grok—generated detailed proposals for terrorist attacks in the United States when prompted by investigators.
In a statement released Wednesday, Hassan emphasized the urgency of the issue, noting, “The AI public safety threat is no longer theoretical.” She warned that if Congress fails to act, it is only a matter of time before a malicious actor uses AI to plan and execute a significant attack. One of the first hearings scheduled for the new Congress in January will address these emerging risks and strategies for improving AI safety.
The 61-page GAO report detailed findings from a performance audit conducted between December 2025 and October 2026. Investigators found that within 12 hours of covert testing, the platforms produced viable plans for either using explosive devices to cause mass casualties in U.S. cities or targeting parts of the electrical grid to inflict economic harm. While the GAO noted that the specific information used was often available on the public internet, it highlighted that AI accelerated the process, allowing for plans that might otherwise take weeks or months to be generated in hours.
Specific tests revealed that two platforms provided instructions for constructing and detonating improvised explosive devices (IEDs), while two others suggested using drones or firearms to attack electrical substations. One test involved a platform suggesting the rental of a vehicle and the purchase of supplies along a specific route to a target location, even recommending a specific date and time based on public events likely to draw large crowds. Although one platform initially refused to provide detailed information on constructing an explosive device, investigators were able to circumvent its safety guardrails to obtain the information.
GAO researchers verified the accuracy of these plans by consulting with experts and conducting site visits, noting some discrepancies such as incorrectly placed roads or high walls that would obstruct visibility. In response to the findings, the tech companies involved—Anthropic, Google, OpenAI, and SpaceXAI—largely stated that they are continuing to refine their models to enhance security. Anthropic acknowledged that determined testers could potentially extract harmful content but noted difficulties in verifying the specific GAO findings. Google indicated that the testing was conducted on earlier versions of its models, while OpenAI and SpaceXAI emphasized their ongoing efforts to improve safety with each model release.
Funny how everyone blames the tech but ignores that the raw information was already out there for decades.
Attacking substations seems more likely than bombs. The grid is fragile and poorly protected in many rural areas.
Twelve hours is terrifying. I hope they actually pass laws, not just hold hearings and forget about it later.
Guardrails are always broken. The real issue is that the underlying data was public anyway, AI just made it faster.